Image→3D input rules — read BEFORE any submit

These rules prevent the two most expensive failure classes: garbage geometry from bad reference images, and wrong polycount paths. Every image→3D submit must pass this preflight — whether through the native 3D CLI path (higgsfield generate create, preferred) or the raw fallback API; the same engine reconstructs the mesh either way. (Param behavior re-checked against docs.meshy.ai 2026-06-11.)

Rule 1 — ONE figure per input image (character sheets)

Meshy reconstructs everything in the frame as a single object. A character sheet with 4 views becomes "a statuette of 4 fused figures".

Mandatory preflight on every user-provided reference:

1. Vision check: "how many separate figures/views are in this image?"
2. If more than one figure → crop. Two valid paths:
   • One best view (front or 3/4) → regular image-to-3d.
   • Each view as its own file → multi-image-to-3d (1–4 images) — better geometry, but ONLY if the pose is identical on all panels. Different poses across panels ⇒ fall back to the single-view path.
3. Cropping in the sandbox: use ffmpeg (ffmpeg -i sheet.png -vf "crop=w:h:x:y" view.png) — PIL is not in the system python.

Rule 2 — pose normalization

Add "pose_mode": "t-pose" (or "a-pose"): Meshy straightens the character regardless of the reference pose. This repairs "awkward" references before auto-rig and matches donor skeletons in the local rig-transfer path.

Rule 3 — framing

• Full body in frame, nothing cut off.
• Clean/plain background; no shadows, text, watermarks, logos.
• Limbs visually separated from the torso (arms not pressed to the body) — fused silhouettes produce fused geometry and break auto-rig.
• One character, no props overlapping the silhouette.
• image_enhancement: false only if the exact input style must be preserved.

Rule 4 — low-poly: two MUTUALLY EXCLUSIVE paths

Do not mix them:

Polycount budgets (game-ready guidance):

Valid API range: 100 – 300 000.

Hard rig limit: models > 300 000 faces are rejected by /rigging — if the mesh came out heavier, run POST /openapi/v1/remesh before rigging.

Rule 5 — economy / QC params

• "enable_pbr": false — skip metallic/roughness/normal maps unless the project needs PBR (base color only is cheaper and faster).
• "target_formats": ["glb"] — don't generate unused fbx/obj/usdz/stl; task completes faster.
• "multi_view_thumbnails": true — 4 cardinal-view PNGs (front/right/back/ left) in the task result for vision QC without downloading the GLB.

Payload templates

Game humanoid, exact budget (recommended default):

{  "image_url": "<cropped single-figure reference>",  "pose_mode": "t-pose",  "should_remesh": true,  "topology": "triangle",  "target_polycount": 20000,  "should_texture": true,  "enable_pbr": false,  "target_formats": ["glb"],  "multi_view_thumbnails": true}

Stylized faceted low-poly look:

{  "image_url": "<cropped single-figure reference>",  "model_type": "lowpoly",  "pose_mode": "t-pose",  "should_texture": true,  "enable_pbr": false,  "target_formats": ["glb"],  "multi_view_thumbnails": true}

Multi-view (character sheet split into same-pose panels):

{  "image_urls": ["front.png", "side.png", "back.png"],  "should_remesh": true,  "topology": "triangle",  "target_polycount": 20000,  "should_texture": true,  "target_formats": ["glb"]}

(endpoint: POST /openapi/v1/multi-image-to-3d)

Preflight checklist (run through it verbatim)

• Vision-counted figures in the reference == 1 (or split for multi-image)
• Full body, clean background, no text/watermark/shadows
• pose_mode set for characters that will be rigged
• Low-poly path chosen consciously (look vs budget) — params not mixed
• target_polycount matches asset class table
• target_formats: ["glb"], enable_pbr: false unless PBR required
• Plan stays under rig limit (300k faces) if rigging follows